Frank
af6cd72d02
- MODEL_DIR -> /home/dank/models/ornith-1.5-35b-a3b-fp8 (37 GB, compressed-tensors per-channel FP8, same qwen3_5_moe arch/mass as Huihui so the pool plan carries unchanged) - CHAT_TEMPLATE_FILE -> scripts/chat-template.jinja (Qwen-Sharp, shared with frank-default-large); the checkpoint's stock template mis-serializes reasoning per its own model card - env/serve/unit headers updated; loader-risk + BF16-does-not-fit note
74 lines
3.6 KiB
Bash
74 lines
3.6 KiB
Bash
# Tunables for frank-default-small (ornith-ai/Ornith-1.5-35B-A3B-FP8,
|
|
# qwen3_5_moe hybrid MoE, Dankserver, RTX 3090).
|
|
#
|
|
# Runtime: FreeToken v0.1.2 (edge-native MoE serving engine).
|
|
# Installed via uv into /home/dank/freetoken/.venv (CPython 3.12).
|
|
# FreeToken keeps expert weights in host RAM with an LRU GPU cache;
|
|
# KV is paged with radix prefix sharing. Only the full-attention
|
|
# layers page KV (~20 KiB/token); the GDN layers use constant-size
|
|
# per-sequence state (mamba slot pool).
|
|
#
|
|
# Model: ornith-ai/Ornith-1.5-35B-A3B-FP8 (HuggingFace safetensors,
|
|
# compressed-tensors FP8, ~35 GB on disk). Ornith-1.5 is a
|
|
# self-improvement RL-trained coding/agentic model built on Qwen3.5 +
|
|
# Gemma4; ~35B total / ~3B active (A3B), same qwen3_5_moe arch and the
|
|
# same ~35 GB FP8 weight mass as the Huihui FP8 it replaces, so the
|
|
# FreeToken memory/pool plan carries unchanged. Native 262144 context.
|
|
# Reasoning model: each assistant turn opens with a <think> block,
|
|
# surfaced separately via the qwen3 reasoning parser (REASONING=on).
|
|
# The RTX 3090 has no FP8 tensor cores, so FP8 runs via upcast (as
|
|
# Huihui did). Replaced batsclamp/Huihui-Qwen3.6-35B-A3B-abliterated-
|
|
# FP8 on 2026-09-01.
|
|
#
|
|
# RISK: this is compressed-tensors PER-CHANNEL FP8 (weights strategy
|
|
# "channel"), the one FP8 layout the runbook flags as a FreeToken v0.1.2
|
|
# loader bug. Load is being verified live; 37 GB fits RAM so a failure is
|
|
# a clean crash-loop, not a swap-thrash. The BF16 sibling
|
|
# (ornith-ai/Ornith-1.5-35B-A3B, ~70 GB) does NOT fit this 62 GB box —
|
|
# never fall back to it (that is the 2026-09-01 Thomson outage mode).
|
|
#
|
|
# Chat template: the checkpoint's stock Qwen template is WRONG for this
|
|
# model (its own model card says so and ships a modified
|
|
# chat_template.jinja for the agentic benchmarks). The fleet standard is
|
|
# the Qwen-Sharp template (scripts/chat-template.jinja, repo-tracked and
|
|
# shared with frank-default-large) — see CHAT_TEMPLATE_FILE below.
|
|
#
|
|
# Pool sizing: GPU budget (~17.6 GiB after weights + fixed caches) is
|
|
# split MoE-first. MOE_CACHE_SIZE pins resident expert slots; NUM_TOKENS
|
|
# pins the KV pool. At 2048 slots / 453005 tokens the split is ~20%
|
|
# residency and ~8.6 GiB KV. Live resize via
|
|
# `ft ctl cache --kv N --moe M` can grow KV by shrinking MoE without
|
|
# restart (the reverse requires a restart).
|
|
#
|
|
# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE.
|
|
# The ROCm HIP runtime reads those bare env vars (see the 2026-08-13 TEMP
|
|
# incident). This box is NVIDIA/CUDA, but the rule keeps the env portable.
|
|
|
|
# Model directory (safetensors checkpoint, ~35 GB, compressed-tensors FP8).
|
|
MODEL_DIR=/home/dank/models/ornith-1.5-35b-a3b-fp8
|
|
|
|
# Chat template: the fleet Qwen-Sharp template (repo-tracked), shared with
|
|
# frank-default-large. The checkpoint's stock template mis-serializes
|
|
# reasoning for this model; the Sharp template renders the <think> trail and
|
|
# the Qwen3.5-family tool-call shape correctly.
|
|
CHAT_TEMPLATE_FILE=/home/dank/serverAppConfigs/scripts/chat-template.jinja
|
|
|
|
# FreeToken venv (uv-managed CPython 3.12, ft CLI, flashinfer, torch+cu130).
|
|
FT_VENV=/home/dank/freetoken/.venv
|
|
CUDA_HOME_DIR=/home/dank/freetoken/.venv/lib/python3.12/site-packages/nvidia/cu13
|
|
|
|
ALIAS=frank-default-small
|
|
BIND_ADDR=0.0.0.0
|
|
LISTEN_PORT=7770
|
|
|
|
# Expert residency: 2048 of 10240 total expert instances (20%).
|
|
# Each slot ~3 MiB → ~6.0 GiB VRAM for the MoE cache.
|
|
MOE_CACHE_SIZE=2048
|
|
|
|
# KV pool: 453005 tokens (~8.6 GiB at ~20 KiB/token).
|
|
# Covers 4 concurrent sessions at ~113k unique context each
|
|
# (more with radix prefix sharing for common prefixes).
|
|
NUM_TOKENS=453005
|
|
|
|
# Max concurrent decoding requests (also caps mamba slot allocation).
|
|
MAX_RUNNING_REQUESTS=4 |